Ontology

Originally a philosophy term about "the system of what exists," but in IT it means organizing the concepts of a field and their relationships into a form computers can work with. It's a blueprint that links concepts as dots (nodes) and relationships as lines — like "OpenAI —builds→ GPT" or "GPT —powers→ ChatGPT."

When this blueprint is filled with actual data, it's called a Knowledge Graph. The info box that pops up on the right when you search for a person on Google is a classic example — instead of reading documents, it pulls answers from a pre-organized web of relationships.

There's a reason this is getting attention again in the LLM era. Language models sound plausible but can get facts wrong (hallucination), while ontologies are the opposite — since relationships are explicit, they don't get things wrong. So there are ongoing attempts to combine "the fluency of LLMs with the accuracy of knowledge graphs." METAL LAB also runs an ontology that places articles not as standalone pieces but on top of company, model, and paper nodes — the "Articles mentioning this term" in this very glossary, and the related-article links at the bottom of each article, are built on top of it.

"We combined it with a knowledge graph to catch hallucinations" — a common line in articles about enterprise AI adoption, referring to a structure that pairs a fluent LLM with an accurate relationship network.

1. Try telling a chatbot: "Take these six as nodes and map out their relationships with arrows: OpenAI, GPT, ChatGPT, Microsoft, NVIDIA, GPU."
2. You'll get a map with relationship labels like "builds," "powers," "invests in," "supplies" — what you just made is a mini knowledge graph.
3. One step further: paste in any AI news article and ask, "Extract the companies, products, and people in this article as nodes, and their relationships as edges." Once the article starts to look like a piece sitting on top of a relationship network — that's exactly why media outlets and companies build ontologies.